Pentagon pressures Anthropic to loosen Claude safeguards
Defense contract leverage meets insider leak economy, Ex-L3Harris unit boss jailed for selling exploits to Russian broker
Images
Dario Amodei and Pete Hegseth. Composite: Getty Images
theguardian.com
Image Credits:Bryce Durbin / TechCrunch
techcrunch.com
Lorenzo Franceschi-Bicchierai
techcrunch.com
A graphic depiction of open laptop overlaying Red Square in Moscow and the logo of Russia's foreign intelligence service SVR.
techcrunch.com
A building featuring Ivanti's logo in red on the top of the building.
techcrunch.com
An illustration of mailbox icons falling out of the cloud with phones in the red background, symbolizing phone spyware.
techcrunch.com
US defense secretary Pete Hegseth met Anthropic chief executive Dario Amodei on Tuesday as the Pentagon pressed for fewer restrictions on how it can use the company’s Claude model, according to the Guardian and Axios. Hegseth reportedly gave Anthropic until Friday to accept the Department of Defense’s terms or face penalties, including the threat of losing a contract that can be worth up to $200m under the Pentagon’s AI procurement program.
The dispute is not about whether the military will use commercial AI—Claude is already integrated into government systems—but about what the vendor will refuse to do once the model is inside classified workflows. Anthropic has marketed itself as unusually “safety-forward” and has resisted uses such as mass surveillance and fully autonomous weapons, the Guardian reports. The Pentagon has signaled it can make that resistance expensive by labeling a supplier a “supply chain risk”, a designation that can freeze a firm out of sensitive procurement and push other agencies to follow.
The pressure campaign lands as the government’s own supply chain for offensive cyber capability shows how leaky “trusted” channels can be. TechCrunch reports that Peter Williams, former general manager of Trenchant, a hacking-tools unit inside defense contractor L3Harris, was sentenced to seven years in prison for stealing and selling seven trade secrets to a Russian broker. Prosecutors said the tools could have enabled access to “millions of computers and devices around the world”; Williams admitted making about $1.3m in cryptocurrency from sales between 2022 and 2025.
Together, the two stories sketch the same institutional pattern from opposite ends. In one, the Pentagon wants a commercial model to behave less like a consumer product and more like a weapon system, with guardrails “tuned for our use cases”, as the department’s chief technology officer Emil Michael put it. In the other, a contractor’s most sensitive capabilities—zero-days and surveillance tooling—were treated as portable personal assets by a senior insider, then monetized through an intermediary that the US Treasury has now sanctioned.
The market signal to AI firms is straightforward: compliance buys access, resistance invites retaliation, and the government can widen the definition of “risk” to include contractual friction. The signal to contractors is equally clear: the most valuable technologies are the ones that move fastest—and the easiest way to move them is through the people already cleared to handle them.
By Friday, Anthropic will either rewrite Claude’s boundaries for a military customer or learn what “supply chain risk” means in practice. On Tuesday, a former hacking-tools executive learned what it means when the same system decides the leak has become too visible to ignore.